Merge branch 'main' of github.com:loreloc/sos-npcs into main

april-tools · Jul 21, 2024 · 11ec3fe · 11ec3fe
2 parents 35e4669 + ad62bb2
commit 11ec3fe
Show file tree

Hide file tree

Showing 6 changed files with 20 additions and 20 deletions.
diff --git a/econfigs/complex-squared-npcs.json b/econfigs/complex-squared-npcs.json
@@ -17,7 +17,7 @@
       "batch-size": 512,
       "optimizer": "Adam",
       "verbose": true,
-      "learning-rate": [5e-4, 1e-3, 5e-3, 1e-2],
+      "learning-rate": [1e-4, 5e-4, 1e-3, 5e-3],
       "num-components": 1
     },
     "models": {
@@ -59,17 +59,17 @@
       },
       "miniboone": {
         "MPC": {
-          "num-units": 64,
+          "num-units": 128,
           "num-input-units": 64
         },
         "SOS": {
           "real": {
-            "num-units": 64,
+            "num-units": 128,
             "num-input-units": 64,
             "complex": false
           },
           "complex": {
-            "num-units": 32,
+            "num-units": 64,
             "num-input-units": 64,
             "complex": true
           }

diff --git a/requirements.txt b/requirements.txt
@@ -1,6 +1,4 @@
 git+ssh://git@github.com/april-tools/cirkit.git@complex-circuits
-torch~=2.2.0
-numpy<2.0.0
 pandas<2.0.0
 scikit-learn
 zuko

diff --git a/src/layers.py b/src/layers.py
@@ -24,7 +24,11 @@ def __init__(
         assert weight1.num_folds == num_folds
         assert weight2.num_folds == num_folds
         super().__init__(
-            num_input_units, num_output_units, arity=1, num_folds=num_folds, semiring=semiring
+            num_input_units,
+            num_output_units,
+            arity=1,
+            num_folds=num_folds,
+            semiring=semiring,
         )
         self._in_shape = (weight1.shape[1], weight2.shape[1])
         self.weight1 = weight1

diff --git a/src/models.py b/src/models.py
@@ -19,8 +19,7 @@
 from cirkit.symbolic.parameters import (
     ExpParameter,
     Parameter,
-    ClampParameter,
-    TensorParameter,
+    TensorParameter, ScaledSigmoidParameter,
 )
 from cirkit.templates.region_graph import (
     RegionGraph,
@@ -184,7 +183,7 @@ def __init__(
             backend="torch", semiring="complex-lse-sum", fold=True, optimize=True
         )
         # Use a different optimization rule for the dense-kronecker pattern
-        self._pipeline._compiler._optimization_registry['layer_shatter'].add_rule(
+        self._pipeline._compiler._optimization_registry["layer_shatter"].add_rule(
             apply_dense_product, signature=DenseKroneckerPattern
         )
         self._circuit, self._int_sq_circuit = self._build_circuits(
@@ -332,7 +331,7 @@ def categorical_layer_factory(
             num_channels,
             num_categories=input_layer_kwargs["num_categories"],
             logits_factory=lambda shape: Parameter.from_leaf(
-                TensorParameter(*shape, initializer=NormalInitializer(0.0, 3e-1))
+                TensorParameter(*shape, initializer=NormalInitializer(0.0, 1e-1))
             ),
         )
 
@@ -347,9 +346,8 @@ def gaussian_layer_factory(
                 TensorParameter(*shape, initializer=NormalInitializer(0.0, 1.0))
             ),
             stddev_factory=lambda shape: Parameter.from_sequence(
-                TensorParameter(*shape, initializer=NormalInitializer(0.0, 3e-1)),
-                ExpParameter(shape),
-                ClampParameter(shape, vmin=1e-5),
+                TensorParameter(*shape, initializer=NormalInitializer(0.0, 1e-1)),
+                ScaledSigmoidParameter(shape, vmin=1e-5, vmax=1.0),
             ),
         )
 
@@ -367,7 +365,7 @@ def dense_layer_factory(
             num_output_units,
             weight_factory=lambda shape: Parameter.from_unary(
                 ExpParameter(shape),
-                TensorParameter(*shape, initializer=NormalInitializer(0.0, 3e-1)),
+                TensorParameter(*shape, initializer=NormalInitializer(0.0, 1e-1)),
             ),
         )
 
@@ -413,7 +411,7 @@ def categorical_layer_factory(
             num_channels,
             num_categories=input_layer_kwargs["num_categories"],
             logits_factory=lambda shape: Parameter.from_leaf(
-                TensorParameter(*shape, initializer=NormalInitializer(0.0, 3e-1))
+                TensorParameter(*shape, initializer=NormalInitializer(0.0, 1e-1))
             ),
         )
 
@@ -428,9 +426,8 @@ def gaussian_layer_factory(
                 TensorParameter(*shape, initializer=NormalInitializer(0.0, 1.0))
             ),
             stddev_factory=lambda shape: Parameter.from_sequence(
-                TensorParameter(*shape, initializer=NormalInitializer(0.0, 3e-1)),
-                ExpParameter(shape),
-                ClampParameter(shape, vmin=1e-5),
+                TensorParameter(*shape, initializer=NormalInitializer(0.0, 1e-1)),
+                ScaledSigmoidParameter(shape, vmin=1e-5, vmax=1.0)
             ),
         )
 

diff --git a/src/scripts/experiment.py b/src/scripts/experiment.py
@@ -224,7 +224,7 @@
 )
 parser.add_argument(
     "--dtype",
-    default="float64",
+    default="float32",
     choices=["float32", "float64"],
     help="The default Torch dtype to use",
 )

diff --git a/src/scripts/utils.py b/src/scripts/utils.py
@@ -111,6 +111,7 @@ def set_global_seed(seed: int, is_deterministic=True):
         if is_deterministic is True:
             torch.backends.cudnn.deterministic = True
             torch.backends.cudnn.benchmark = False
+    torch._dynamo.config.cache_size_limit = 24
 
 
 def get_git_revision_hash() -> str: